The Review of Economic Studies LTD

The Review of Economic Studies Ltd.
Authority and Communication in Organizations

Author(s): Wouter Dessein
Source: The Review of Economic Studies, Vol. 69, No. 4 (Oct., 2002), pp. 811-838
Published by: The Review of Economic Studies Ltd.
Stable URL: http://www.jstor.org/stable/1556723
Accessed: 25/05/2009 01:31
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=resl.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.
The Review of Economic Studies Ltd. is collaborating with JSTOR to digitize, preserve and extend access to
The Review of Economic Studies.
http://www.jstor.org
Review of Economic Studies (2002) 69, 811-838 0034-6527/02/00320811$02.00
? 2002 The Review of Economic Studies Limited
Authority and in
Communication
Organizations
WOUTER DESSEIN
Universityof Chicago and CEPR
First version received October 1999; final version accepted January2002 (Eds.)
This paperstudies delegation as an alternativeto communication.We show that a principalprefers

to delegate control to a better informedagent ratherthan to communicatewith this agent as long as the
incentiveconflict is not too large relativeto the principal'suncertaintyaboutthe environment.We further
identify cases in which the principal optimally delegates control to an "intermediary",and show that
keeping a veto-righttypically reduces the expected utility of the principalunless the incentive conflict is
extreme.
1. INTRODUCTION
This paperis concernedwith the old saying that "knowledgeis power". In organizations,much
of the informationused in decision makingis dispersedin the hierarchy.Lower-levelmanagers,
for example, are often much better informed about consumer needs, competitive pressures,
specialized technologies or marketopportunitiesthan their superiors.The financialpress is full
of stories about how companies have pushed decision rights lower in the hierarchyin orderto
profit from this local knowledge.1 For the same reason, newly acquiredsubsidiariesare often
left with substantialautonomy.The goal of this paperis to betterunderstandwhy an uninformed
principal(the companyowners,seniormanagement)may grantformaldecision rightsto an agent
(senior or middle management)who is betterinformedbuthas differentobjectives.We arguethat
a principaloften delegates authorityin orderto avoid the noisy communication,and hence the
loss of information,which stems from these differencesin objectives.
At firstsight, it may seem a puzzle why keeping authorityandletting the agentreportwould
not always weakly dominatedelegation.By keeping authority,the principalhas always the option
to rubberstampthe proposals of the agent, but she may also refrain him from implementing
projects which are obviously not in the interest of the organization.By delegating authority,
in contrast,the principalcommits to never reverse the agent's decisions. We will nevertheless
argue that delegation is typically a better instrumentto use the local knowledge of the agent
than communication.Key to our analysis is that differencesin objectivesbetween principaland
agent are often systematic and predictable.It is, for example, well documentedthat managers
may be short-termbiased, status-quobiased, risk-averse, empire builders etc. Whenever the
principaland the agent systematicallydisagree on a certainaction dimension,the principalwill
not rubberstampa naive recommendationby the agent of his preferredaction, but try to correct
for the "bias"in objectives.As the agent is not naive but anticipatesthis, communicationis then
inherentlystrategicand-in equilibrium-noisy. Hence, the centraltrade-offin our paperis one
between a loss of controlunderdelegationand a loss of informationundercommunication.
1. Among firms decentralizingdecision rights in the 1990's are AT&T,GeneralElectric, EastmanKodak,Fiat,

Motorola,United Technologies,Xerox and, recently,Ford.
811
812 REVIEWOF ECONOMICSTUDIES
Model. In order to analyse this trade-off, we develop a stylized model in which the
principal (she) must screen among a range of projects which differ from each other on one
dimension. The agent (he) has superiorinformationon which project is best for the principal,
but his objectives differ in a systematic way. He could, for example, always prefer a larger
project than the principal(size-bias). For simplicity, this bias is constant and positive. Section
3 provides a discussion of the kind of biases we have in mind. The privateinformationof the
agent is assumedto be soft, thatis the agent cannotproveor certify his knowledge.Furthermore,
following GrossmanandHart(1986) andHartandMoore (1990), we posit thatprojects(actions)
cannotbe contracteduponand,hence, the principalcannotuse a standardmechanismto elicit the
privateinformationof the agent. The principal,however,can contracton the authorityover the
project.Indeed, to engage in a project,some critical resourcesare needed which are controlled
by the principal.This implies thatthe agent normallyneeds the fiat of the principalto implement
a project, but the principalcan also delegate decision rights to the agent by grantinghim the
authorityover the use of these criticalresources.
In our organization,the principalthus faces the choice betweenfully delegatinga task to a
betterinformedagent or to order the latter what to do after having consulted him. If she keeps
decision rights and consults the agent, a game of strategiccommunicationtakes place in which
each equilibriumis characterizedby a partitionof all possible states of natureand where the
agent introducesnoise into his signal by only specifying to which partitionelement the realized
stateof naturebelongs. Given the informationprovidedby the agent,the principalthen takes the
action which maximizes her expectedutility.Such a strategicinformationtransmissionhas been
firstanalysedby Crawfordand Sobel (1982), hereafterreferredto as CS. While communication
always involves a loss of information as long a preferences are not perfectly congruent, a
centralresult of their paper is that the closer the preferencesof agent and principal,the better
is communication.The loss of informationeven goes asymptoticallyto zero when differencesin
objectives disappear.Delegation, in contrast,results in a loss of control since the agent always
takes a decision which is biased relativeto the firstbest. Similarly,this loss of controlbecomes
smaller when the agent's preferencesare closer to those of the principaland disappearsin the
limit. At firstsight, the optimalallocationof authorityis thus not trivial.
Results. Our main finding is that the principal optimally delegates control as long as
the divergence in preferences is not too large relative to the principal's uncertaintyabout
the environment.Thus, if the state of natureis uniformly distributed-a standardassumption
in almost any applicationof the Crawford-Sobel model-the principal prefers delegation to
communicationwhenever the agent's bias is such that informativecommunicationis feasible.
The larger the uncertaintyabout the environment,the larger is the range of biases for which
the principaldelegates control. More generally,for any given informationstructure,delegation
dominatescommunicationif the bias is sufficientlysmall. Indeed, for any continuousand twice
differentiabledistribution,as the agent's bias tends to zero, a principalwho keeps control and
communicates,will take an action which is on average an infinite times furtheraway from the
firstbest thanthe action the agent would take.
The intuitionbehind these results lies in the natureof the "screening"mechanismat work.
For the agent to be induced to tell the truth,it is typically necessary that his messages become
increasingly noisy as they recommend actions which go furtherin the direction of his bias.
Intuitively,if an empire-builderrecommendsa "large"project,this message is less informative
than if he recommendsa "small"project. In addition, to preventthe agent from exaggerating
his information,the increase in noise in subsequentsignals should also be proportionalto his
bias. As a result, the betteris communication,and thus the more messages the agent is able to
send, the largermust be the averagenoise in these messages relative to the bias. In terms of the
DESSEIN AUTHORITYAND COMMUNICATIONIN ORGANIZATIONS 813
CS-equilibrium,the finerthe equilibriumpartitionbecomes in absoluteterms,the more coarse it

is relativeto the bias. Hence, the more informativeis communication,the betterit is to delegate
authorityto the agent and avoid communication.
In contrast, if we keep the bias constant, but change the information structure,
communicationdominates delegation if the uncertaintyabout the environmentis sufficiently
small. Indeed, while changing the information structuredoes not affect the loss of control
under delegation, a more precise prior allows the principal to select an action which is on
averagemuch closer to the firstbest, and thus substantiallyreducesthe loss of informationunder
communication.While informativecommunicationmay then dominatedelegation(e.g. when the
priorof the principalis very "steep"andb is not too largerelativeto the support),communication
is then typically very noisy. Simulationswith (truncated)normaldistributionsand quadraticloss
functionsshow thatdelegationis optimalunless the bias is so largethatcommunicationis almost
uninformative:regardlessof the varianceof the distribution,only if the bias is such thatthe agent
recommendsthe same actionin more than98%of all statesof nature,communicationdoes better
thandelegation.
To conclude, we investigate whether the principalcan improve upon the pure delegation
outcome by some limitedforms of delegation:
- In Section 6, we show that in the leading example of Crawfordand Sobel,2 for moderate
biases, the principaloptimallydelegates decision rights to an intermediarywith objectives
in between hers and her agents'. Doing this, the principal may prevent the agent
from implementingextreme projects, withoutjeopardizingtoo much the communication
concerning small and intermediateprojects. Delegation to the agent remains optimal for
small biases.
- In Section 7, finally,we considerdelegationwith veto-powerfor the principal,a mechanism
which is known as the "closed rule" in Political Science (see Section 2). Again, one
might conjecturethatdelegatingbut retainingveto-powershould at least weakly dominate
complete delegation.For reasonablechoices of the status quo and a uniformdistribution,
however, we show that keeping veto-power is only beneficial for large divergences
of preferences. For small or moderate biases, keeping veto-power results in additional
variancein the deviationfrom the first-best,and complete delegationis optimal.
Outline. The paper is organized as follows: Section 2 gives an overview of the related
literature.Section 3 describes the model. Section 4 characterizesthe equilibrium for given
decision rights. Section 5 then analyses the optimal allocation of authority.We go on to
investigatethe value of delegatingcontrolto an intermediary(Section 6) andthe value of keeping
veto-power (Section 7). Section 8, finally, discusses various extensions of our model: profit-
sharingarrangements,privateinformationconcerningthe agent's bias, repeatedinteraction,and
verifiableinformation.We conclude in Section 9.
2. RELATEDLITERATURE
The incentive view on delegation. So far, the economic literatureon organizationshas
emphasizedan incentive based rationalefor delegation.In particular,Aghion and Tirole (1997),
show that a principalmay delegate formal authorityto an agent in orderto give the latterbetter
2. In the leading example of CS, principaland agent have a quadraticloss function and the state of natureis
assumedto be uniformon [0,1]. So far, it has been the workinghorse for almost any applicationof CS.
incentives to acquire information.3While the focus of Aghion and Tirole is on the impact of
authorityon the informationstructure,we take the informationstructureas given-the agent is
assumed to be better informed-and we investigatehow the allocation of authorityaffects the
use of this privateinformation,providinga purelyinformationalrationalefor delegation.
The informational theory of legislative rules. Both the incentivebasedrationaleandour

purely informationalrationalefor delegationhave a counterpartin the "informationaltheory of
legislative rules" in political science. In a very influentialpaper,Gilligan and Krehbiel(1987)
adopt the leading example of CS4 in order to provide a rationale for the use of restrictive
amendmentproceduresemployed in the U.S. House of Representatives.In particular,Gilligan
and Krehbiel (GK) is concerned with the motivation of the "closed rule", under which an
uninformedlegislaturecan only veto but not amend a proposal of a committee. If a proposal
is vetoed, the status quo (which is taken exogenous) prevails. As Aghion and Tirole, G K
assumes that this committeemust make an effort to become informed.For the leading example
of Crawfordand Sobel, GK finds that when the preferencedivergencebetween the legislator
and the committeeis small, the "closedrule"is to be preferredover the "openrule"underwhich
the legislaturecan freely amend the proposal of the committee. While GK mainly emphasize
the closed rule's impacton incentivesto acquireinformation,5Krishnaand Morgan(2000) have
recently shown that even in the absence of an informationacquisitionproblem,the closed rule
dominatesthe open rule as long as informativecommunicationis possible underthis open rule, a
resultwhich is similarto ours.While delegationis relatedto the "closedrule",6Section 7 shows
that in the model of GK, for reasonablechoices of the statusquo, delegationstrictlydominates
the closed rule unless preferencesdivergencesare extreme.In this sense, our paperpoints to a
missing elementin the reasoningof G K, which shouldexplain why the committeethen does not
receive full decision power.
Other related literatures. Finally, our paper differs from a number of literatureson
"informationrevelation"in its assumptions on the commitment ability of the principal. By
adopting an incomplete contractingapproach,we clearly departfrom the standardprincipal-
agent model in which the principalelicits privateinformationby designing a mechanism.While
this approach may explain many institutions, we feel that the underlying premise that the
principalcan perfectlyand withoutcost commit herself to any mechanismis too strongin many
organizationalcontexts.
In sharp contrast with the mechanism design literature, the literature on strategic
communicationor "cheaptalk",initiatedby CS assumes that no commitmentat all is possible.
When we think about communicationin organizations,this is also ratherunrealistic.Indeed,
it is an insight of the propertyrights literaturethat ownership or more generally the control
over critical resourcesconfers authorityto its holder.Hence, the principalcan commit to never
overrulethe decisions of the agent by delegatingto him the controlover these criticalresources.
We discuss some of the recent literatureon cheap talk in Section 8.2.7 Most closely related is
3. As arguedby Szalay (2000), however,it may be optimalto limit the discretionof the agent, and force him to
choose between extremeoptions in orderto provideeven strongerincentivesfor informationacquisition.
4. See footnote 2.
5. On p. 325, for example, they state "The dominantfocus of the paper has been on the informationalrole
of committees, more specifically on a committee's incentive to acquire expertise which, in equilibrium,may be used
beneficially."
6. If the status quo were sufficiently extreme, the closed rule would even be identical to delegation. In the
mentioned papers, however, this is never the case, since they assume that the status quo policy may be optimal. As
is clear from Proposition4, the closed rule then resultsin a very differentoutcome thanpure delegation.
7. Otherrecentpapersare Banerjeeand Somanathan(2001), Battaglini(2001), and Krishnaand Morgan(2001).
Unlike CS, these papersconsidermore thanone agent.
de Garidel-Thoronand Ottaviani(2000), which also comparesdelegationwith communication,

and investigateswhat happenswhen the informationis noisy, the directionof the agent's bias is
uncertainand the strategicsophisticationof the principallimited.
Finally, a few studies, such as Holmstrom (1984), Melumad and Shibano (1991), and
Armstrong(1994), posit that the principalcan commit to a decision rule but not to monetary
transfers.As noted by Holmstr6m,such a decision rule only limits the discretionof the agent to
a subset of actions and thus boils down to a partialform of delegation.In contrastto our paper,
which stresses the loss of informationwhich occurs when a commitmentto (full) delegation
fails, Holmstromand Armstrongderive some generalpropertiesof the optimalpartialdelegation
while Melumadand Shibanoemphasizethatthe sender(the agent) does not always benefit from
communicationand hence may try to avoid it. The main difference with our paper,however,
is that these paperspresumethat the principalcan reverse some actions (which she determines
ex ante) and at the same time is able to commit neverto reverseothers.
3. THE MODEL
A (profitor non-profit)organizationhas the opportunityto engage in a valuableproject.There
are an infinityof potentialprojects,but only one projectcan be undertaken.While an agent (he)
is hired to implementthis project,a principal(she) initially controlsthe criticalresourcesof the
organizationwhich are needed to initiate any of these projects.The principalcan be the CEO or
the ownerof a firm,but in principle,any hierarchicalrelationshipin an organizationcould fit our
model.
Preferences. Projects differ from each other on one dimension and can be represented
by a real numbery e R. (Alternatively,projectsmay have differentdimensions,but agent and
principal agree on all but one dimension.) With each project y is associated a monetarygain
and/orprivatebenefit Up = Up(y, m) for the principaland a privatebenefit UA = UA(y, m, b)
for the agent, where m is a randomvariableand b a parameterof dissonancebetween agent and
principal.The utility of the principalreachesa uniquemaximumfor y = m and can be rewritten
as
Up(y, m) = Up(m, m) - e(ly - ml)
where e"(.) > 0 and e'(0) = 0. Similarly,the utility of the agent is maximized for y = m + b
and can be rewrittenas
UA(y, m) = UA(m+ b, m) - A(IY - (m + b))

where ge (.) > 0 and g' (0) = 0. Wlog, we assume b > 0. We will often referto b as the bias of
the agent.
Systematic biases in agency relationshipsare well documented. It is well accepted, for
instance, that managershave a propensityto cause their department,division or firm to grow
beyond the optimal size, i.e. they are empire builders and undertaketoo many investments.8
They further seldom take externalities on future managers into account and, hence, are
excessively oriented towards short-termprofitabilityand results. Employees, concerned with
theircareerperspectives,will favourprojectswith a high visibility or a close contactwith senior
management;they may, for the same reason, prefer projects which allow the acquisition or
improvementof importantskills or avoidriskyprojects.Managersalso often internalizetoo much
the interestsof their subordinates.Bertrandand Mullainathan(1999), e.g. provideevidence that
managershave a preferencefor paying high wages. In the same vein, managerswith close ties to
8. See, for example, Jensen (1986).
their personnelmay fire too few employees duringa restructuring.Anecdotalevidence of other

biases abound:employees are claimed to be effort-averse,status-quobiased, etc.
It is worthnotingthatbiases often ariseendogenouslyas the productof inherentlyimperfect
incentive schemes. Division managers' salaries depend in general on the performanceof their
division, which distortsincentives if projectsinvolve externalitieson other divisions. Similarly,
shareholdersmay partiallycontrol managerialshort-termismby the way remunerationdepends
on reportedearnings and changes in stock-marketvalue. In order to simplify the analysis,
however,we will treatthe bias as exogenous in the core of the paper.In Section 8.1, though,we
briefly discuss the impact of differentincentive-schemes,such as profit-sharingarrangements,
which aim to reducethe agent'sbias.
Information structure. Only the agent observes the random variable m whose twice
differentiabledistributionfunction F(m), with density f(m), is supportedon [-L, L]. The
other parametersof the utility functions are common knowledge. Though not made explicit
in the model, the superior information from the agent can be seen as an externality from
implementingactions in previous periods or from his "proximity"to the business environment
(clients, suppliers,competitors).In the core of the paper, we assume that informationis soft,
thatis the agentcannotcertify or "prove"his information.No restrictionis imposed on the set of
messages which can be sent by the agent.In Section 8.3, we discuss how relaxingthis assumption
may affect our results.
Authority and contracts. We adoptan incompletecontractingapproachby positing that

projects (actions) cannot be contracted upon and that to engage in a project, some critical
resourcesareneeded, which are initiallycontrolledby the principal.Resourceswhich we have in
mind, are (a) the assets of the organization,(b) the name of the organizationand more generally
the right to contracton behalf of the organizationwith thirdparties,and, (c) to some extent, the
human resourcesof the organization.9While projectscannot be contractedupon, the principal
may then grantsubordinatesauthorityover the use of the resourcesneededto initiatea project.10
This can be done by contracts,job-descriptions,corporatecharters,customs or, in the extreme
case, by selling some of the assets of the organizationto the agent."1If an agenthas controlover
the criticalresources,he can initiate a projectwithoutassistance of the principal:he has formal
decision rights.If the principalkeeps authority,on the otherhand, the agent needs her fiat. This
fiat can take the form of some "signatures",may requiresome concreteactions by the principal,
or may imply that the principal(or her staff) takes full care of the initiationstage. In any case,
the principalthen fully controlsthe projectchoice. Once a projectis initiated,it still needs to be
implementedby the agent, but cannotbe reversedany more.
The timing is as follows. (i) The principal decides whether or not to delegate the agent
authorityover the use of the criticalresources.(ii) The agent learnsm and initiatesher preferred
project if she has authority.If the principalhas not delegated authority,she may ask the agent
to make a recommendation,and then initiates a project.(iii) The agent implementsthe project.
9. This control stems from the fact that the principalcan contractwith employees, that is hire, promote,demote
and fire them.
10. By arguing that authoritynot only stems from ownership of assets, we follow several recent articles. For
example, Aghion and Tirole (1997) arguethat"Authoritymay more generallyresultfrom an explicit or implicit contract
allocatingthe rightto decide on specifiedmattersto a memberor groupof membersof the organization"(p. 2). Rajanand
Zingales (1998) stress access, definedas the ability to use, or work with, a criticalresource,as an alternativemechanism
to allocate power. Baker, Gibbons and Murphy(1999) investigate how decision rights may be delegated informally
throughself-enforcingrelationalcontracts.
11. Furthermore,the owner may delegate the authorityto delegate some of these decisions to employees at lower
levels in the organization,and so forth.
We assume thatcontrolrightsover resourcescan only be allocatedto the agent at the initial date
and, hence, are always unconditional.
4. DELEGATIONVS. COMMUNICATION
In our model, the principal has two instrumentsto use the local information of the agent:
delegation and communication. In this section, we study both instruments separately and
comparetheir comparativestatics.
4.1. Delegation
Suppose first that principal and agent cannot communicate with each other. The principal
then delegates control or takes an uninformedaction. If the agent has control, he implements
y = m + b, yielding
Up(m + b, m) = Up(m, m) - t(b)
to the principal.Hence, the principaldelegates authorityif and only if b is smallerthan a cut-off
value b' given by
L
(b) =_miny L (ly - ml)dF(m)-
-L
Obviously,the principal'sexpected utility increasesas the agent'sbias decreasesand reachesthe
firstbest as b tends to zero.
4.2. Communication(Crawford-Sobel)
Consider now the polar case where the principal cannot commit to let the agent decide, but
communicationbetween agent and principalis feasible. Since projectsare non-contractible,the
principalthen always undertakesthe projecty which maximizes her expected utility conditional
on her beliefs uponm. Hence, the only thing communicationmay achieve,is changingthe beliefs
of the principal.This form of communicationis often referredto as "cheaptalk" and was first
studiedby Crawfordand Sobel (1982). In CS, a better-informedsender (the agent) may reveal
some of his informationby sendinga possibly noisy signal to a receiver(the principal),who then
takes an action (initiates a project) which determinesthe welfare of both. The only constraint
on the information transmission is that the agent's message and the principal's subsequent
decision form a Bayesian equilibrium.Formally,an equilibriumis characterizedby (i) a family
of signalling rules q(nlm) for the agent, where for every m E [-L, L], q(nlm) is the conditional
probabilityof sending message n given state m, and (ii) a decision rule y(n) for the principal,
where y(.) is a mappingfrom the set of feasible signals N to the set of actions 7Z,such that
- for each m E [-L, L], if n* is in the supportof q(. m), then n* maximizes the expected
utility of the agent given the principal'sdecision rule y(.),
- for each n, y(n) maximizes the expected utility of the principal,taking into account the
agent's signalling strategyand the signal she receives in orderto updateher prior of the
distributionof m.
As shown by CS, all equilibriain this communicationgame are characterizedby a partition

of [-L, L], where the sender (the agent) introducesnoise into his signal by only specifying
to which partitionelement the realized state of nature belongs. As their model encompasses
ours, the following proposition,which is a variantof Theorem 1 in CS, can be shown to hold.
Let a = (ao, ..., aN) denote a partitionof [-L, L] with N steps and dividing points between
steps ao, . .,aN, where -L = ao < al < . * < aN = L. Define, for all a, a e [-L, L], a < a,
y(a, ) -- argmax f Up(y, m)f(m)dm.
Proposition 1 (Crawford and Sobel: Communication Equilibrium). If b > 0, then

there exists a positive integer N(b) such that,for every N with 1 < N < N(b), there exists
at least one equilibrium(y(n), q(n\m)), where
q(y(ai- ,ai) m) = 1 if m E (ai-1,ai), (1)
UA(y(ai, ai+1), ai) - UA(y(ai- 1, ai), ai) = 0, (A)

(i = 1,...N-1),
y(n) = y(ai-1,ai) if n E (ai-1,ai) (2)
ao =-L, and aN = L. (3)
Further,all other equilibriahave relationshipsbetween m and the principal's inducedchoice of
y that are the same as those in this class for some value of N with 1 < N < N(b); they are
thereforeeconomicallyequivalent.12
Proof. The proof follows directlyfrom CS, Theorem 1. 11
In equilibrium,only a finite number N < N(b) of actions are thus implemented with
positive probability,and the states of naturefor which each of these actions is best for the agent
form an interval,and these intervalsform a partitionof [-L, L]. The partitiona is determined
by (A), a well-definedsecond-orderlineardifferenceequationin the ai's, and (3), its initial and
terminalconditions.Equation(A) is an "arbitrage" conditionwhich says thatfor statesof natures
that fall on the boundariesbetween steps, the agent is indifferentbetween the associatedvalues
of y. Given our assumptionsabout UA, this condition is necessary and sufficientfor the agent's
signallingrule to be a best responseto y(n). Similarly,(2) gives a best responseof the principal
to the signallingrule (1).
Equilibrium selection. Although there is, in general, a multiplicity of economically

differentequilibria,CS provide some sufficientconditions underwhich the expected utility of
both the principaland the agent are ex ante (that is, before the agent knows the state of nature)
maximizedby the same equilibrium.13 Under these conditions, this is the equilibriumwith the
largestnumber,N(b), of partitionelements. As CS, we thinkthatit is reasonablefor the players
to coordinateon this (ex ante) Pareto-superiorequilibrium.Doing this, we allow communication
to be as powerfulas possible.
Comparative statics. Underthe same conditionsas above, CS establisha sense in which

communicationimproves when the receiver (the principal) and the sender (the agent) have
more similar preferences.They show that for a given b, the principal always strictly prefers
equilibriumpartitions with more elements and that the largest possible number of partition
elements N(b) weakly decreaseswith b. In the limit, as preferencesof senderand receivertend
12. CS Theorem 1, for instance,proposesthatq(n I m) is uniform,supportedon [ai_ , ai] if m E (ai_-, ai).
13. While these conditions are quite stringent,they are always satisfiedwhen F(m) is uniformand UP and UA
dependon y and m only throughy - m, as in our model.
to coincide, communicationeven becomes perfect: in the leading example of CS, the noise in
the sender'smessage then tends to zero in the most informativeequilibrium,a resultwhich holds
for any distributionF(m), as shown by Spector(2000). In contrast,once preferencesdivergeby
more than a given finite amount,only uninformativecommunicationis consistent with rational
behaviour.
A first comparison. From the previous analysis, the comparativestatics of delegation

and communication are very similar. When the agent's bias is very large, no informative
communicationis possible, but also delegationis suboptimal:the principalthen optimally takes
an uninformeddecision. When the agent'sbias decreases,communicationimproves,from which
CS conclude "directcommunicationis more likely to play an importantrole, the more closely
related are agent's goals". However, also the expected utility under delegation increases as b
decreasesand both implementthe firstbest as b tends to zero. At firstglance, it is thus not clear
how the optimal allocationof authoritywill vary with b.
5. THE OPTIMALALLOCATIONOF AUTHORITY

We are now ready to endogenize the allocation of formal decision rights. As there exists in
generalno tractablesolution to the second-orderlineardifferenceequation(A) of Proposition1,
unless F(m) is uniformlydistributed,we firstfocus on thatsimple case. We subsequentlydiscuss
to what extent our resultscarryover to more generaldistributions.
5.1. Uniformdistribution
From CS, we know that all communicationequilibriaare fully characterizedby a partitionof
[-L, L], where the agent tells the principalto which partitionelement the state of nature m
belongs. As a measure for the minimum loss of informationunder communication,it will be
useful to define the minimalaverage size of the partition elements,denotedby A(b):
A(b) -=2L/N(b)
where N(b) is the maximumnumberof partitionelements in equilibriumgiven b. Note thatthis
measure underestimatesthe real loss of informationif partitionelements are unequal in size.
Similarly,a measurefor the loss of controlunderdelegationis given by the bias b.
Denoting by y* = y(m) the action undertakenby the principal under communication,a
sufficientconditionfor delegationto be strictlypreferredover communicationis that
E(ly*- ml) > b. (4)
Even when the equalityholds, the principalstrictlyprefersto delegatebecause of the variancein
1y*- ml. Since F(m) is uniform,we have that
E(ly* - m) > A(b)/4 (5)
where the inequalityis strictif and only if partitionelements are unequalin size.14
From (4) and (5), a sufficientconditionfor delegationis thus A(b)/b > 4. In the remainder
of this section, we show that A(b)/b > 4 holds whenever N(b) > 3. In addition,while A(b)
goes to zero as b tends to zero, A(b)/b then tends to infinity.We conclude by showing that also
for N(b) = 2, we necessarily must have that E(ly* - ml) > b, and thus wheneverinformative
communicationis possible, the principalprefersto delegate authorityand avoid communication.
14. If partitionelementsdifferin size, m is more likely to belong to largerpartitionelementsfor which E(ly* -m |)
is larger.
ao Y1 al (Y1+ Y2)/2 Y2
FIGURE 1
a2 - a = -
ao + 4b
Key to our results will be the observationthatpartition elements are increasinglylarge as

we move up the interval,where this increaseis proportionalto the bias. Suppose that Yl and Y2
are two adjacentactionswhich aretakenwith a positive probabilityin equilibriumand denote by
ao, al anda2 the dividingpoints of the equilibriumpartitionrespectivelyprecedingyi, following
yi and following Y2.
Since the principalis not restrictedin her project choice, she always initiates the project
which is equal to the averagestate of natureof a partitionelement, thatis Yl = (ao + al)/2 and
Y2 = (al + a2)/2. At the dividing point m = al, the agent must be indifferentbetween yi and
Y2.As a result, we also have that al = (Yi + y2)/2 - b. As can easily be seen on Figure 1, the
lattertwo conditionsimply that
a2 - al = al - ao + 4b.
Lemma 1. The size of a partition element is always 4b larger than the size of the
preceding one:
ai+l - ai = al - ao + 4ib, i = 1,..., N(b) - 1.
Lemma 1 explains us neatly how the "screening mechanism" of the communication

equilibriumworks: the agent is induced to tell the truthby the fact that "larger"messages are
more noisy, which makes "exaggerating"his privateinformationvery costly. The largerthe bias
of the agent, the largerthe incentive to exaggerateand, hence, the largershould be the increase
in noise to keep the agent on his toes.
Note that the above result is in line with some straightforwardintuitions on real world
communication.Commonsense tells us thatif a personwith a preferencetowardslarge projects
recommendsa "large"project,then his message is not very informative.If, on the other hand,
the same person proposesa "small"project,this is much more revealing.Intuitively,one would
thus expect thatthe more the agent's message pleads for an action which goes in the directionof
his bias, the noisier this message will be, or in the terms of our model, the largerthe size of the
associatedpartitionelement. On the otherhand,if the agent had no bias at all, therewould be no
reasonfor a "large"message to be more noisy thana "small"message. Hence, the increasein the
noisiness shouldbe somehow proportionalto the bias.
Clearly, Lemma 1 seriously limits the effectiveness of communication as a means of
information aggregation relative to delegation. Indeed, good communication requires a lot
of messages or partitionelements, where each partitionelement must be 4b larger than the
precedingone. Hence, if communicationis very good in absolute terms, the average size of a
partitionelement, A(b), will be very large relative to the bias b. Concretely,from Lemma 1, it
follows that
1 N(b)-1
A(b) = al - ao + N(b) . "= 4ib > 2[N(b) - ll]b. (6)
As a result, the larger is N(b) and thus the better is communication,the worse communica-
tion performs relative to delegation. When b tends to 0, A(b)/b even goes to infinity. More
importantfor our purpose, from (6), whenever N(b) > 3, we have that A(b) > 4b and thus
E(ly* - ml) > b.
We concludeby showingthatalso for N(b) = 2, the principalprefersto delegateratherthan
to communicate.Denote Al = al - ao and A2 = a2 - al. FromLemma 1, then A2 = A1 + 4b
so that if the principalkeeps authority,she implements y = ao + A /2 if m E (ao, al) and
y = al + Al/2 + 2b if m e (al, a2). Given that F(m) is uniform, this implies again that
E (ly* - m 1) > b, where the equalityholds if and only if AI = 0. The next propositionstatesour
firstmain result:
Proposition 2. If F(m) is uniformly distributed over [-L, L], the principal prefers
delegation to communicationwheneverb is such that informativecommunicationis feasible.
The following corollaryis an immediateconsequenceof Proposition2:
Corollary 1. If F(m) is uniformly distributed over [-L, L], the principal delegates
control rights to the agent if and only if b < b', where b' is such that the principal is indifferent
betweenan uninformeddecision and a biased decision:
1 lL
t(b') = -
| t(m)dm. (7)
L Jo
Corollary1 providesa firstdeterminantfor the allocationof authority:delegationis optimal

when preferencesbetween agent and principalare not too far apart.Two other determinants,
which stem from (7), are that:
- Delegation is more likely when the amount of private informationof the agent is large.
A measure for the informationaladvantageof the agent is given by the variance of m,
a2 = L2/3. From (7), b' increases with a2 and goes to infinity as a2 goes to infinity.
The principalthus delegatescontrolas long as the bias of the agent is not too largerelative
to the amountof privateinformationof the agent. Given that no communicationoccurs in
equilibrium,this result is very intuitive:an increase in the variancedecreases the pay-off
of an uninformeddecision while it has no impact on the pay-off of a delegated decision.
Withoutthe priorknowledge of Proposition2, however,this would be less straightforward
since more uncertaintyalso induces more communication:from Lemma 1, N(b) increases
weakly with L and also informativecommunicationis possible for a largerrangeof values
of b as L and thus increases.
- Delegation is more likely when the principal is more risk-averse.If the principalkeeps
authority,she takeson averagean unbiasedaction,butthe deviationfromthe optimalaction
has a large variance.Hence, the more concave her utility function, the more attractiveis
the constantbias which prevailsunderdelegation.
The leading example of Crawford-Sobel. An interestingimplication of the previous

analysis is that limb,o E(ly* - ml)/b = oo or in the limit, the principalis an infinite number
of times furtheraway from the firstbest undercommunicationthanunderdelegation.This could
stem fromthe fact thatthe principal'sloss relativeto the firstbest is only of second-orderin b with
delegation, while there is afirst-orderutility loss with communication.We verify this as well as
our no communicationresultin case of the leading exampleof Crawford-Sobelwhich allows for
a closed-formsolution.Suppose F(m) is uniformlydistributedon [0, 1], Up (y, m) = -(y-m)2

and UA(y, m) --(y-(m+b))2. As shownby CS, Section4, the expectedutilityof the principal
undercommunicationis then given by
1 b2(N2- 1)
^EU^p= 12N2 3
where N is the numberof partitionelements.As shownby CS, EUp is maximizedfor N = N(b)
which is given by the smallest integergreateror equal to
1 1 2\1/2
2 2( b)
One can verify immediately that for N > 2, EUp < -b2 and thus delegation is optimal
whenever informativecommunicationis possible. We now investigate what happens when b
becomes small. In the limit, as b goes to zero, N(b) _ 1//2b. Hence
b aEUp
limbo0 EUp =-- and =-1/3,
3 ab b=0O
which implies that communicationleads to a first-order loss in b. In contrast,delegation only
resultsin a second-orderloss: the principal'sutility is then given by -b2 andhence aEUIp1b=0=
2b = 0.
5.2. Generaldistributions
We now discuss to what extent the results and intuitions obtained for a uniform distribution
can be generalized to other distributions.When the agent's bias is small, we show that for
any distributionF(m), communicationperforms very badly compared to delegation, just as
with a uniformdistribution.For large biases, in contrast,we give a sufficientcondition on the
"informativeness"or "steepness"of F(m) such that very noisy but informativecommunication
does betterthan delegation.Obviously,this raises the question how large is then the parameter
rangefor which delegationis optimal.To give an idea on this, we finallyreportsimulationresults
for (truncated)normaldistributionsanda quadraticloss function,which show thatonly for biases
for which communicationis extremelynoisy, the principalkeeps control.For small or moderate
biases, the principaldelegates authorityto the agent.
Small biases. For b small, the principalalways delegates control to the agent. Indeed,
when the agent's bias is small, communication must be very informative in order to
dominate delegation. When communication is very informative, however, there must be a
large numberof partitionelements, whose size-just as with a uniformdistribution-must be
increasinglylarge as we move up the intervalin orderto refrainthe agent from exaggeratinghis
privateinformation.As a result, the averagesize of the partitionelements will be huge relative
to the bias, which yields our second main result:
Proposition 3. Consider the most informativecommunicationequilibriumgiven b. For

any F(m), in the limitas b tends to zero, a principal who keeps controland communicates,is on
average an infinitetimesfarther awayfrom m than a principal who has delegatedcontrol:
E(lY* -m) )
limb-->0 = co. (8)
We providea sketchof the argument,a formalproof is given in the Appendix:

If F(m) is not uniform,condition (A) of Proposition1 implies that the increase in size of
adjacentpartitionelements will be larger (smaller) than 4b if the density of m is downwards
(upwards)sloping, as y(ai-1, ai) is then smaller(larger)than (ai + ai-1)/2.
For communicationto have any chance against delegation when b is small, however, the
size of the partitionelements must tend to zero as b tends to zero.15But in the lattercase, the
increasein size will be approximately4b, regardlessof the distributionF(m). Intuitively,when a
partitionelementbecomes small, for any distributionF(m), the differencein density f (m) along
the partitionelement becomes negligible relativeto the averagedensity of the partitionelement,
just as with a uniform distribution.Condition (A) of Proposition 1 then implies that partition
elements increase at a rate of 4b as long as b is of the same orderof magnitudeas the partition
elements.
Indeed, let (ai-1, ai) and (ai, ai+1) be two adjacentpartitionelements, then it is shown in
the Appendixthatif (ai-1, ai) and (ai, ai+l) are small
'
ai+1 - ai (ai - ai-1) + 4b - [(ai - ai-i)2 + (ai+l - ai)2] ( f (ai)
where we have neglected all terms (and only those) that are in the third or higher order of
(ai - ai-1) and (ai+l - ai). As a result, as long as b is of the same order of magnitude as
the partitionelements, these partitionelements increasedramaticallyas we move up the interval,
which can be used to show (8). Partitionelements may stop to increasewhen b - (ai - ai-I)2,
but then (ai - ai-1)/b ~ 1/Vb, which also implies (8).
Large biases. For a sufficiently large bias, it is obvious that delegation is inferior to
communication,as the principal'spay-off with the latter is bounded by what she can achieve
if she took the decision in ignorance. If we only consider biases b for which informative
communicationis feasible (N(b) > 2), though, then with a uniform distribution,delegation
always dominatescommunication.The uniformdistribution,however,is the limit case of how
agnostic the principalcan be concerningwhich action she optimally should take. If we fix the
bias and the support[-L, L], and let F(m) become more informative,then communicationwill
dominate delegation when the uncertaintyabout the environmentis sufficientlysmall. Indeed,
while changingthe distributionof m does not affect the loss of controlunderdelegation,a more
precise priorallows the principalto select an action which is on averagemuch closer to the first
best, and thus substantiallyreducesthe loss of informationundercommunication.
In order to formalize the above intuition, we only consider symmetric single-peaked
distributions, which has the advantage that the range of biases for which informative
communicationis feasible is independentof F(m).16
Lemma 2. If F(m) is symmetric,then communicationis informative(N(b) > 2) if and

only if b < L/2.
Proof. See the Appendix. II
Proposition 4. Assume F(m) is symmetric and b < L/2, then informative

communicationdominatesdelegation if F(m) is such that
f g(lml)dF(m) < g(b). (9)

L
15. If this is not the case, it is shown straightforwardly(see the Appendix)that (8) holds.
16. In a previousversion we also consideredasymmetricdistributions.The resultsare not qualitativelydifferent.
f(m)' ,
-L al L
FIGURE 2
Given b and F(m), communicationis informativeand dominatesdelegation
Proof. If (9) holds, the principalobtainsa higherutility by takingan uninformeddecision

(y* = 0) than by delegatingcontrolto the agent. Since communicationcannot make her worse
off, she then optimally keeps controls. Moreover,as b < L/2, informativecommunicationis
then feasible. 11
Corollary 2. Assume F(m) symmetricand b < L/2. If theprincipal has a quadraticloss

function, then informativecommunicationdominatesdelegation if a (m) < b.
Proposition4 and its corollary tell us that, for a given bias b, keeping control is optimal
if F(m) is sufficiently informative. If b < L/2, informative communication then occurs
in equilibrium.This contrasts with Proposition 2 which states that, for a given distribution
F(m), delegating control is optimal if b is sufficientlysmall. Takentogether,this suggests that
delegationdominatescommunicationif the bias of the agent is small relativeto his informational
advantage,just as we found for a uniformdistribution.
Figure 2 gives an example of a bias b and a steep single-peakeddistributionF(m) for
which informativecommunicationis feasible (the best partitionequilibriumis (-L, al, L)) and
keeping controlis optimal.
Note thatin the above example, communicationis not very informativeand if the bias were
smaller(e.g. equal to b/2), delegationwould again be optimal.The following simulationresults
suggest thattypically only very noisy communicationmay dominatedelegation.
Normal distributions. How largeis the parameterrangefor which delegationis optimal?

To have a quantitativeidea on this, we have performednumerical simulations for a class of
truncated normal distributions,17 and a quadratic loss function, (I\y- m ) - (y - m)2. We
find that the principaldelegates authorityif and only if b < b'a, where b' = 0.955, and this
independentlyof the variance(we checked for a = 0.5, a = 1 and cr = 2). For biases larger
or equal to b'a, communicationoccurs in equilibrium,but the agent recommends the same
"high"action in at least 98.6% of all states of nature.The outcome is thus very close to the
"no communicationresult"of the Uniform distribution:only if communicationis very noisy, it
beats delegation.
17. Concretely,we have assumedthatm - Af(0, ar)/cte, supportedon [-4cr, 4cr]where cte = Jf4. dA(O, ra)
0-99994. The results,however,are insensitiveto the truncationpoints as long as L > 4a, with [-L, L] the support.
6. THE VALUEOF AN INTERMEDIARY

In practice, one often observes that top management delegates control to intermediaries-
supervisors,managers, external consultants-with objectives between the principal's and the
agent's. We investigate thereforewhether, in our setting, the principalmay have an incentive
to delegate control to a middleman with a bias bi e (0, b), ratherthan to the agent. In the
latter scenario,the middleman-which is assumedto be equally uninformedas the principal-
takes a decision afterhaving consultedthe agent. Obviously,such an allocationof controlfosters
communicationcomparedto the case where the principalis in charge, and limits the loss of
controlcomparedto full delegationto the agent.18
Given that the only role of the intermediaryis to communicatewith the agent and take a
decision, it seems reasonablethat the principaldisposes of a pool of candidateintermediaries,
who all differ in their biases. An obvious candidateis the intermediarymanagement.Product-
line managers, for example, typically have interests in between those of top management
and the individual product managers (say, due to the presence of implicit incentive schemes
or relation-specificinvestments).19Secondly, entities which are external to the organization,
such as consultants,may act as intermediariesin matterswhere both agent and principalhave
vested interests.Finally, intermediariesmay also be createdin the marketplace.In the software
industry,for example, open source companies, such a Collab.net,act as intermediariesbetween
independentopen source programmersand commercial corporations,who release the source
code of new technologies to these intermediaries.20To simplify the exposition, we will make
the extreme assumptionthat for any k E (0, 1), the principalcan find an intermediarywith bias
bj = (1 - k)b. This standsin contrastwith the agent, who is uniquein being informedand hence
cannotbe replaced.We furtherassumethatthe intermediaryis not allowed to delegateon his turn
the authorityto the agent. Since the intermediarywill have an incentive to do this, the principal
then must clearly define the decision rights.21
The questionwe wantto answeris thus: "giventhe agent's bias, if theprincipal couldfreely
choose an intermediary'slevel of bias, would she choose a biased intermediary?".To keep the
problemtractable,we only characterizethe optimal delegation scheme for the leading example
of Crawfordand Sobel. We subsequentlydiscuss in a more general way the logic behind the
value of an intermediary.
Proposition 5. If F(m) is uniformlydistributedon [O,1] and theprincipal has quadratic

preferences, that is t(y - m) = (y - m)2, then
* For b < 1/6, the principal delegates authorityto the agent.

* For b E ]1/6, V/2/4[, the principal delegates authorityto an intermediarywith bias b* E
(0, b).
* For b > v/2/4, there is neitherdelegationnor communication.
18. A relatedexplanationfor the role of intermediariesin organizationsis given by Mitusch and Strausz(1999),
which show how an intermediarycan improve things in a communicationgame where the principalcannot commit to
a decision rule. In their model, the intermediarydoes not receive control, but is simply a "mediator"which is more
congruentwith the agent than the principal.Several papershave furthershown (see, for example, Vickers (1985)) that
firmswhich are engaged in competitiveinteractionsoften have an incentiveto strategicallydelegatecontrolto a manager
who is not a profit-maximizer.
19. Note that top managementcould potentially increase (decrease) the congruence between middle and lower
managementby reducing(increasing)the span of controlof middle management.
20. See Lernerand Tirole (2000).
21. This is realistic:it is common practicein organizationsthatthe signatureof a certainperson (of a certainrank)
is obligatoryto have access to critical resourcesof the firm. Hence, authoritycannot be passed on without the approval
of top management.
FIGURE 3
Delegation decision in leading example Crawfordand Sobel (1982)
Proof. See the Appendix. 11
Figure 3 summarizesthe optimalallocationof authoritywith and withoutan intermediary:

If the choice of intermediariesis limited, one may want to know for what values of b and k
the principalwould use an intermediary.Proposition5 gives a partialanswerto this by providing
a necessary condition: only if the "optimal"bias bI belongs to (0, b), an intermediarywith
bi : bI may be useful. As an example, suppose that bi = b/2 exogenously. One can verify
thatthe principaldelegatescontrolto this intermediarywhenever 2- < b < .
is
Why delegating control to an intermediary valuable for intermediatebiases? Basically,
delegation to an intermediaryis a commitmentdevice. From Proposition 1, communication
between agent and principal is a noisy signalling game in which a finite number of actions
{Yl,.. . , YN}are implementedwith positive probability.Since in this signallingequilibrium,the
optimalresponseof the principalis always a purestrategy,the agentknows exactly which action
will be implementedfor which signal. The agentthuspossesses a limitedform of discretion:he is
able to implementany action y e Y -{ ...., YN}. By delegatingauthorityto the intermediary,
the principalchanges the set of actions Y and thus the discretionof the agent. For example, if
b = L/2 and F(m) is uniformon [-L, L], then Y = {0} if the principalkeeps control, since
no informativecommunicationis then possible. In contrast,if the principaldelegates control to
an intermediarywith bias bl = b/2 = L/4, we have that Y = {-L/2, L/2}.22 In this sense, it
is interestingto investigatewhat is the "optimalamountof discretion"to be given to the agent,
where the set of actions Y does not have to satisfy the conditionsof Proposition1. Previouswork
has alreadystudiedthis questionand the following resultholds:23
Lemma 3. If F(m) is uniformlydistributedon [-L, L] the optimalamountof discretion

to be given to the agent, consists of putting an upper bound y' on the actions which the agent
may implement,wherey' = L - b if b < L and y' = 0 if b > L.
Proof See Holmstr6m(1977, Section 2.3.2). 11
22. Indeed, if the intermediarycommunicateswith the agent, then at most two partitionelements are possible,
where the size of the second partitionelement equals the size of the first plus 4bf = 2b. Hence, the agent sends a
low message if m e (-L, -b) and a high message if m e (-b, L). The intermediarythen implementsrespectively
b
Yl- 2 +2 and Y2 = L2b + .
23. See Holmstr6m(1977, 1984), Melumadand Shibano(1991), and Armstrong(1994).
Intuitively,the principal may reduce the average deviation from the optimal project by
puttingan upperbound on the projectswhich the agent is allowed to undertake.Furtherlimiting
the discretionof the agent,however,only harmsthe principalif her prioris uniformlydistributed.
While she may reduce the bias in some states of nature,this will always lead to an increase in
the bias for other states of nature,generatinga mean-preservingspreadof the deviation which
is betteravoided given her concave preferences.In our example where b = L/2, optimally,the
principalwould allow the agent to choose any action y < L/2. Note thatthis upperboundwould
be binding with a probability1/2.
In practice, however, it may be difficult to commit to such a policy, since the principal
will also be tempted to intervene when the agent wants to implement small or medium-sized
projects. Here, an intermediarymay play an importantrole. By delegating decision rights
to an intermediary,the principal may prevent these very large projects from being initiated
without jeopardizing the communicationconcerning small or medium-sized projects. Indeed,
for b = L/2, if an intermediarywith bias bi = b/2 = L/4 has control, the agent has discretion
on {-L/2, L/2}. Hence, by delegating control to this intermediary,the principal effectively
implementsthe optimal upperboundL - b, while the agent still has the option to communicate
that a smaller project is optimal. The expected deviation from m, E(ly* - m I), is then 3b. In
contrast,if the principalkeeps control or delegates to the agent, the expected deviationfrom m
equals b. Unless the loss function of the principalis very convex, the principalthus prefers to
delegate to the intermediary.
It is finally intuitive that an intermediaryis only useful for intermediatepreference
divergences, as shown in Proposition 5. For intermediatebiases, the benefits of constraining
projectsto be smallerthanL - b, are substantialandcommunicationperformsnot so bad relative
to delegation.In contrast,when preferencesof principaland agent are close, the probabilitythat
the optimal upper bound L - b is binding, is very low. The benefits of limiting the agent's
discretion are then small. At the same time, communicationperforms very badly relative to
delegation.
7. DELEGATIONWITHVETO-POWEROR "CLOSEDRULE"
As anotherform of limited delegation,the principalmay delegatea decision to the agentbutkeep
the rightto veto this decision ex post, in which case the statusquo prevails.A CEO or a division
manager,for example,may be delegatedthe taskto come up with a planto redesigna product-line
or to restructurethe firm or a division, but theses plans need a final "thumbsup" of the boardof
directorsor top management.Importantly,delegationwith veto-poweris also a regularpractice
in some political organizations.In the U.S. House of Representatives,for example, under the
closed rule, a parliamentarycommitteehas the right to proposea bill which can only be vetoed,
but not amended.
In this section, we comparethe effectivenessof delegationwith veto-power,or equivalently
the "closed rule",with pure delegation.As we discussed in Section 2, the "closed rule" and its
raison d'etre has been the object of intense debate and study in Political Science. In a seminal
paper,Gilligan and Krehbiel(1987) have used the leading example of CS to show thatfor small
biases, the closed rule yields a higher expected utility to the House than the "open rule"under
which the House can freely amend the proposals of the committee and which is equivalentto
communication.24In this section, we show that unless preferencedivergences are very large,
puredelegationdoes even betterthanthis closed rule. In case of the leading example of CS and a
24. By focusing on a more efficient equilibrium,Krishnaand Morgan(2000) even show that the closed rule does
betterthancommunicationwheneverthe latteris informative,that is as long as b < 1/4.
statusquo equalto the mean of the distribution,delegationbeatsthe closed rule unless preference
divergencesare so extremethatno informativecommunicationis possible.
For simplicity and following the traditionof the literatureon legislative rules, we assume
that F(m) is uniformlydistributed.In accordancewith GK, we makethe following assumptions.
There is an exogenously given status quo Yo e [-L, L]. The sequence of moves under veto-
based delegation is the following. First, the agent learns m. Secondly, the agent proposes a
project y e Y to the principal.Finally, the principalchooses between y and yo. Similar to the
CS communicationgame, the veto-delegationgame has typically several equilibria.We will
focus on the one proposedin Krishnaand Morgan (2000, Proposition8) which is thus far the
most efficient equilibriumidentified in the literature.The following propositioncharacterizes
this Krishna-Morganequilibrium:
Lemma 4. Underdelegation with veto-power(closed rule):
(i) If b < (L- yo)/2:

- the agentproposes to implement:
* y=m+bifm<yo-borm>yo+3b.
* Y = Yo if m E [yo - b, Yo].
* y= yO+2bifm e (yo, y +2b].

* y=y + 4bifm e (yo + 2b,yo + 3b].
- the principal vetoes all projects y e (yo, yo + 4b)\{yo + 2b}, and rubberstampsall
otherprojects.
(ii) If b > (L- yo)/2:

the agentproposes y = m + b if m < yo - b and y = Yootherwise,
- the principal vetoes y if and only if y > yo.
Proof See the Appendix. 11
We comparethe outcome underpure delegationwith the above equilibrium.Given that the

leading example of CS is the traditionalworking horse in the literatureon legislative rules,
we first make the comparison under the specific assumptions of this leading example. For
concreteness,we also assume thatthe statusquo equals the expected optimal value of y, that is
Yo= E(m). We subsequentlygeneralizeour resultsto more generalloss functionsand supports,
and to differentstatusquos.
Proposition 6. Consider the leading example of CS, in which F(m) is uniformly

distributedon [0, 1] and theprincipal has a quadraticlossfunction. If yo E(m) = ?, thenpure
delegationstrictlydominatesdelegation with veto-power(closed rule) if and only if b < 1/4.
Thus, for the leading example of CS and yo = E(m), we obtain the result that pure
delegationdominatesthe closed rule unless the bias of the agent is so large that no informative
communicationbetween agent and principalis possible (indeed for b > 1/4, N(b) = 1). In
contrast,for b > 1/4, the closed rule is optimal. Given that the status quo equals the expected
value of m, the closed rule then also weakly dominateskeeping control.
What is the intuitionbehind these results?First considerthe case where b < (1 - Yo)/3 =
1/6. From Lemma 4 the closed rule (veto-power) then introduces a variance in bias of the
implementedproject, but does not reduce the average bias.25Since deviationsfrom the firstbest
are increasinglycostly, the principalthen strictlypreferspuredelegation.Secondly, considerthe
case where b > 1/6. FromLemma4, the closed rule then still introducesa considerablevariance
in the bias, but on average,this bias is smaller than b.26 Thus,for b > 1/6, the principal must
trade-offa smaller average deviation with a larger variance under the closed rule. As long as
1/6 < b < 1/4, the reductionin the averagebias is not sufficientto compensatefor the large
variance,and the principalpreferspure delegation.In contrast,for b > 1/4, the reductionin the
average bias is sufficientlylarge and veto-power becomes optimal. Interestingly,for b > 1/4,
the agent is allowed to implement any project y < yo = 1/2 under the closed rule. Keeping
veto-poweris then a tool to refrainthe agentfrom implementinglarge projects, while allowing
him discretion concerning small projects. Whereasfor small biases, such an upperboundon y
would be far too tough, for large biases, it approachesthe optimalupperboundL - b.27
We now generalizethe above result for any convex loss function,any support[-L, L], and
any statusquo yo E [-L, L]. Define b(yo) - (L - yo)/3:
Proposition 7.
(i) If b < b(yo), pure delegation strictlydominatesdelegationwith veto-power(closed rule).

(ii) If b > b(yo), 3f(x) such that delegation with veto-power (closed rule) dominatespure
delegation if the loss function of the principal, t(x), is less convex in the sense of Arrow-
Pratt than t(x).
The intuitionfor Proposition7 is identicalto the one of Proposition6. For b < (L - yo)/3,
keeping veto-poweronly introducesvariancein the bias, but does not reduce the averagebias.
In contrast,for b > (L - yo)/3, the principalhas to trade-offa largervariancewith a smaller
averagebias. Whetheror not the closed rule is optimal,dependsthen on the risk-aversionof the
principal.
Note that b(yo) is decreasingin yo, suggesting thathaving veto-poweris more useful when
the statusquo is larger(or more in the directionof the agent's bias). Intuitively,from Lemma 3,
it is only optimal to refrain the agent from implementinglarge projects. If the status quo is
large, veto-powergives the principalan effective tool to do this withoutjeopardizingthe agent's
flexibilityto implementsmall and medium-sizedprojects.28For the leading example of CS, one
can easily show thatthe cut-off value for b above which the closed rule is optimal,is decreasing
in yo. While for Yo = 0, the closed rule is never optimal,29for yo = 1, the closed rule always
dominatesdelegation.
25. Indeed, for m E [1/2, 1/2 + 2b], the expected deviation from m ranges from 0 to 2b. Similarly,while for m
e [1/2- b, 1/21, the expected deviationis b/2, for m e [1/2 + 2b, 1/2 + 3b], it equals 3b/2.
26. Reason is that m is then less likely to belong to the range [1/2 + 2b, 1/2 + 3b] where the bias in the
implementedprojectis largerthanb.
27. See Lemma 3.
28. If the principal could commit ex ante to a status quo point, he would choose y* such that the optimal
upperboundcharacterizedin Lemma 3 is implemented,that is yo = L - b. FromLemma4, the veto-delegationgame is
then equivalentto the optimal partialdelegationscheme in which the agent can pick any action y < L - b.
29. For b < V/3/6, puredelegationis then best while for b > V3/6, the principaloptimallykeeps full controland
implementsy = 1/2.
8. EXTENSIONSAND DISCUSSION
To conclude, we discuss some important extensions of our basic set-up and their likely
implications.
8.1. Extendingthe space of contracts

The main result of our paper is a comparisonof two institutions:communication(cheap talk)
vs. delegation. While these are indeed the only instrumentsavailable to the principal given
our assumptionson the non-contractibilityof projects and profits, one may wonder whether
we have not biased our results in favour of delegation by restrictingthe space of contractsin
this way. In this section, we argue that delegation will typically be a complementratherthan
a substituteto other contractualarrangements,in particularthose institutions, such as profits
sharing,promotions,career concerns etc., whose aim it is to better align the incentives of the
agent with those of the organizationor principal.Indeed, the smaller the bias of the agent, the
better delegationperformsrelative to communication.Therefore,by reducingthe agent's bias,
these mechanismsmay make delegationoptimalwhereasinitially the principalwould have kept
control. Suppose, for example, that pay-offs are in monetaryterms and that the principalis free
to offer simple sharing contractsthat awardto the agent a percentageof the monetarypay-offs
deriving from the project. If we suppose that the agent derives a privatebenefit of a project y
given by
B(y) -KA -kA(m+b_ y)2
and the principalmaximizes a contractibleprofit
Jr(y) --K - k(m _ y)2,
then awardingthe agent a shares of 7r(y), induces the agent to maximize a functionof the form
B'(y)=KA-kA m+ (kA sk)b-
Thus, giving the agent a share s in total profits, reduces its bias from b to b' = kA+skb. The
optimalamountof profitsharings for a given bias will dependon whetheror not ex ante transfers
from agent to principalare possible and on the wealth constraintof the agent. At least with a
uniform distribution,however, the principalwill always delegate control for a largerrange of
biases thanif no sharingcontractswere possible.
8.2. Uncertaintyover the bias and repeatedinteraction

A convenientsimplificationof our model is that the bias of the agent is known to the principal.
Intuitively,though, as long as there remainsa systematicaveragebias in the agent's objectives,
introducinguncertaintywill not greatly affect our results. In a setting where the agent has a
bias b only with a probabilityp, Stocken and Morgan (2000) provide a first analysis of the
communicationequilibriumwhich then prevails.One of their findings is that the informational
efficiency of communicationmay be either smalleror largerthanif the agent's bias were known
to be p*b. In the examplesthey provide,however,the differenceis limited.30Therefore,one may
expect the trade-offbetween communicationand delegationto be only quantitativelyaffected.
A different picture is obtained if the unknown tastes of the agent were symmetrically
distributedaboutthe principal'stastes, thatis the agent is unbiasedin expectations.As shown by
30. Of the orderof 2% or less.
de Garidel-ThoronandOttaviani(2000), if F(m) is uniformlydistributedandthe bias takeseither

a positive or negative value, there exists a communicationequilibriumin which the principal
rubberstampsthe proposalsof the agent as long as these proposals are not too extreme.31This
can be shown to improveupon pure delegation,which suggests that if divergencesin objectives
are randomratherthan systematic,the principaloften optimallykeeps control.
Finally,when the bias of the agent is uncertain,it is also very importantto take into account
the impactof repeatedinteraction,as has become clear from Avery and Meyer (2000). In a two-
period, two-actionmodel where agents have an uncertainbut positive bias,32they show that the
presenceof a second periodacts as a first-perioddisciplinedevice: agentsarethen less biased and
may even show a negativebias in ordernot to be perceivedas a type with a largebias. In a typical
equilibrium,the principalrubberstampsthe recommendationof the agent in the firstperiod, and
follows the advice of the agent in the second period if and only if the latterhas recommended
the "small"projectin the first period. Given the binary natureof the example, this outcome is
identicalto the one in which the principalalways delegates authorityin the firstperiod, and lets
her second period delegationdecision depend on the firstperiod action of the agent. Intuitively,
one may expect similarresults if there is a continuumof projects,where the agent refrainsfrom
implementing"large"projects in an initial period out of fear of being perceived as extremely
biased and hence losing control in a later period. To the extent that this puts a reasonable
upperboundon the projects which are implementedin initial periods, intertemporalincentives
may then greatly improve the efficiency of delegation.33This reputationaldelegation, and its
comparisonwith reputationalcheap talk, are exciting topics for futureresearch.34
8.3. Hard vs. soft information

A crucial assumptionin our analysis is that the private informationof the agent is soft. Our
motivationfor this is thatthe best thing to do typically dependson a wide rangeof factors,which
the principalmay find very hard to foresee, of whom she may not understandtheir relevance
and significance, let alone verify their magnitudeor nature.Therefore,almost any project can
be claimed to be optimal. In some settings, however, it is realistic to assume that the agent
can somehow "prove"or "certify"his local knowledge and that the principalhas the time and
the backgroundto assert this information.This may radically change the trade-off in favour
of communication.As shown by Seidmannand Winter (1997), in the extreme case where the
agent, in any state of nature, can certify his informationat no cost to him or the principal,
communicationis even perfect. Obviously, the principal then never delegates. This suggests
some interestingpredictionson differencesin the hierarchicalstructureof firmsacrossindustries.
Even when privateinformationcan be certifiedor verified,this is often a very time-consuming
process. In fast changing industries,the benefits may then not outweigh the costs in terms of
incurreddelays in the decision making and the heavy burdenthis puts on managerialattention.
Privateinformationis then de facto soft and our model predictsratherflat organizationswith a
decentralizedorganizationstructure.In contrast,when the businessenvironmentis very stableor
information-processingcosts are negligible, one may expect more centralizeddecision making.
31. Concretely,thereexists an equilibriumin which the principalrubberstampsall proposalsy e [-L + b, L -b],

where the bias of the agent is eitherb or -b.
32. Avery and Meyer are concernedwith the applicationwhere a biased evaluatormust recommendwhetheror
not to hire a candidate.In our language,"hire"is the large project,"do not hire",the small project.
33. As pointed out by Avery and Meyer, however, the intertemporalincentives may also be too strong, making
agents far too cautious.
34. Note, however, that this reputationalcheap talk is differentfrom the one studied by Ottavianiand Sorensen
(1999), where the expertis concernedto appearwell informedratherthancongruent.
9. CONCLUDINGREMARKS
This paperhas stressedthe limits of the use of soft informationin organizations.To the extent
thata senior managercannotverify the claims of a betterinformedsubordinate,she is in general
betteroff delegatingdecision rights to this subordinatethan relying on the informationshe can
induce fromhis claims. Intuitively,while the subordinatemay not tell her what she shoulddo, he
will use all his informationwhen he himself takes the decision.
This simple resulthas potentiallyimportantimplicationsfor the design of organizations.It
suggests that centralizationof authorityis only optimal if top managementhas the information
which is importantto the main decisions, or is able to check and verify the information
provided by lower levels of the hierarchy.At first sight, this is in line with the tendency of
firms to focus on core activities, i.e. activities on which they have a profound knowledge,
and to outsource other activities. Similarly, the trend of the last two decades towards more
decentralizationand empowerment,highlighted by the business press, may find its origin in
a rapidly changing business environmentwhich causes the knowledge of top managementto
become quickly obsolete. In response to an increased foreign competition in the 1980's and
1990's, for instance, many large American cooperations(ITT, IBM, GeneralMotors, Eastman
Kodak, and Xerox, ...) have changed the organizationaldesign of their organization and
frequently pushed decision rights lower in the organization.35An incentive based theory of
communicationcosts in hierarchies,where these costs arise endogenously from the necessity
to check and understandreportsprovidedby agents prone to distort their information,will be
needed to throwfurtherlight on these issues.36
APPENDIX
A. 1. Generaldistributions
Small biases
Proof of Proposition 3. Let (ao, al,..., aN(b)) characterizethe most informativepartitionequilibriumgiven

b, let hn be the length of the nth partitionelement, hn - an - an-1, and let h be the largestpartitionelement. The proof
of (8) follows directlyfrom the following threelemmas.
Lemma Al. As h tends to zero,
nn 6f'(an)
hn+l = hn + 4b-[h2 + hn+1 f (an) 6 (A.1)
wherewe have neglectedall terms in the 3-rd or higherorderof hn and hn+ .
Proof. Let us consider the (n + l)-th partitionelement and denote Yn+l - y(an+i, an). We will now express
Yn+l as a function of hn+l and an using Taylorapproximationsin which we neglect all terms (and only these) which
are in the thirdor higherorderof hn+i. For this purpose,Vm e (an, an+l), we approximatee(lm - Yn+l l) and f(m) by
(Im -Yn+l ) = 2t"(0)(m - Yn+l)2

f(m) = f(an) + f'(an)(m -an) + f (an)(m -an)2
where both m - Yn+l I and m - an are fractionsof hn+1.
35. A discussion of some case studieson this can be found, e.g. in Brickley,Smith and Zimmerman(1997).
36. The literatureon communicationand processing costs (Radner(1993), Bolton and Dewatripont(1994), van
Zandt (1999) etc.) has assumed a team theoreticalframeworkwithout incentive problems. Note that the existence of
explicit time/resourcecosts of communication,centralin these papers,would only reinforcethe preferencefor delegation
in our model.
Denoting kn Yn+1 - an, we have that

hn?l
kn = argmink| "(0)(k - x)2[f (an) + f'(an)x + 1 f "(an)x2]dx.
Takingthe FOC yields
hn+l (x kn)[f(an) + f'(an)x + l f "(an)x2]dx = 0
or still
f (an)h h3
2n+l kn f (an)hn+ 1 + 3 n+1- -k+kn f 2n28 h2n01 + f)h4+1-kf(an)
8 n n 6
from which
f(an)/2 + f'(an)hn+1/3 + ft(an)hn+1/
n n+l f (an) + ff(an)hn+l /2 + f t(an)h2n+l/6 _
A first-orderTaylorexpansionin hn+ I of the term in bracketsyields37
1 -
kn = hn+l 2 + hn+l (f '(an)/3)f (an) (f'(an)/2)(f (an)/2)
from which
k_ hn+l h+1 f'(an)
kn
2 + 12 f(an)
Since kn = Yn+l - an and hn+ = an+1 - an, it follows that
1 f
an + an+l + (an+1 - an)2 (an) (A.2)
2 12 f (an)
where we have neglected all terms in the 3-rd or higherorderof (an+l - an).
In the same way, for (an - an- 1) small, one can show that
2 12 f (an)
where we have neglected all terms in the 3-rd or higherorderof (an - an- 1). At m = an, the agent must be indifferent
between Ynand Yn+l from which
Yn+1- (an + b) = an + b - yn.
Substituting(A.2) and (A.3) we find
= an _ an-l + 4b -[(an+l f
an+-an _ an)2 + (an _ an-1)2] 1 (an)
6 f (an) \
or, equivalently,
hn+ =hn +4b + [h2 + h2 f(an)l

]
where we have neglected all termsin h3, h3+1 and higherordersof hn and hn+
Lemma A2. If limb,O h = 0, then

E(ly*b-ml)
limb_0 -- (A.4)
Proof. Let us fix two points xl and x2 of the support of F(m), where -L < xl < X2 < L. Since
[F(x2) - F(xl)] > 0, to prove that (A.4) holds, it is sufficientto show that
limbO( (E(ly*- m| Im (xl, x2)) ) (A.5)
37. As this expansionis multipliedby hn+l, a second-orderTaylorexpansionwould only add additionaltermsin

n+l' which subsequentlywould be neglected.
h3n+,
Abusing notation,let us denote from now on by N(b) the numberof partitionelements fully containedin [x1, x2], and
by hn the size of the n-th partitionelement thatis fully includedin (xl, x2). Defining further
=
fJ(m)
maxmE(xlx2) f(m) '
we have from Lemma Al thathn+i > hn + 4b - /[h2 + h2 1]. We now firstshow that
limbo0 E(hi)/b = oo (A.6)

holds (PartA and B) and then thatthe latterimplies (A.5) (PartC):
Let us denote by q(b) the numberof partitionelements fully includedin [x1, x2] for which /^h2 < b and by Q(b)
the numberof partitionelements for which /h2 > b. We have q(b) + Q(b) = N(b). By assumption,if b goes to zero,
q(b) + Q(b) goes to infinity.We considertwo cases.
(A) First,assume that
lim Q(b) > 0.

N
Denotinglimb,o Q(b)- > 0, we have
E(hi) 1 Q(b)E(hi JIAhi > b) 1

limbo- b l!b. > limbh_o
0 ?.=
b b N
(B) Secondly,assume that
)
limb-o =Q 0. (A.7)
Let us denote by nh(b)the averagelength (numberof partitionelements) of a series of adjacentpartitionelements for
which ih2 < b. For any two adjacentpartitionelements hi and hi+l which both belong to such a series, it follows from
(A.1), that
hi+1 > hi + 2b
and thus
E(hi I| th2 < b) > -1i- ( -il 2ib = [n(b) - l]b. (A.8)
If limb-O0 N(b) = 0 holds, then limb-o0 E(hi) = limbo E(hi I trh2 < b). Moreover, it also follows that
limb^o n(b) = oo.38 From(A.8) it then follows that
limb_o b > limbo[hn(b)- 1] = oo.
(C) We now show that (A.6) implies (A.5). Denoting by fmin = minme(xl,x2) f(m) and by (ai-1, ai) the ith
partitionelement includedin (xl, x2), then
l(ai-l, ai) -mlf(m)dm > a +ai mm idm

i-_l i-l 2
>hi h m fmindm f h
20 2 4
and thus E(ly* - m 1m e (Xl, x2))
> -
F(xl)-F(X2) y(ai-l ai) mf(m)dm
F(aN) -F(ao) ,N fnmin h2

F(xl)- F(x2) i-=1 4 F(aN)-F(ao)
- F(aEhi)
fn
> F(aN) F(ao) 1 hN h, _ fmin F(aN )
4 F(x1)-F(x2) N i=1 I 4 F(x1)-F(x2)
38. Indeed,suppose thatlimb-o n(b) = n* with n* a finite number,then

Q(b) Q(b) 1
limbo N
>N
lim4 Q(b) + n(b)Q(b) I + n*
a contradiction.
where F(xN)-F(x) tends to 1 as h tends to zero. From (A.6) then (A.5) and thus also (A.4). l!
Lemma A3. If limbo h > 0, then limb,O E( ) = o.
Proof. Let us denote by ai the largest partition element of the most informative equilibriumand by ) the
0 0 and limb_,o{miny E(ly -
probabilitythatm e ai.Given the finite supportof F(m), if limb,o h > 0, also limbo (>
mlim E a)} > 0. Hence limb/,o E(ly* - ml) > 0 and limb,.o E(ly* - ml)/b = oo-. 1
Large biases
Proof of Lemma 2. From Proposition 1, whenever a communicationequilibriumexists with more than one
partitionelement, therealways exists a communicationequilibriumwith two partitionelements. Hence, it is sufficientto
show that a partitionequilibriumwith two partitionelements exists if and only if b < L/2. A partitionequilibriumwith
two partitionelements exists if and only if 3a E (-L, L) such that
a+b-y(-L,a)= y(a,L)-a-b q
2b = y(a, L) + y(-L, a) - 2a.
The R.H.S. of this equalityis a continuousdecreasingfunctionof a since both ay(a, L)/aa < 1 and ay(-L, a)/aa < 1.
Moreover,the R.H.S. equals L if a = -L and -L if a = L. Hence, if and only if b < L/2, 3a E (-L, L) such thatthe
equalityholds, and a communicationequilibriumwith two partitionelements exists. II
A.2. Delegation to an intermediary
Proof of Proposition 5. Suppose the principaldelegates control to an intermediarywith bias bj = (1 - k)b.

Denoting by am2 (kb) the residualvarianceof m the intermediaryexpects to have afterhearingthe equilibriumsignal of
the agent, it is easy to verify thatdelegationto the intermediaryyields
EUp = Up(m, m) - a2 (kb) - (1 - k)2b2. (A.9)

This expression reflects the fact that quadraticloss equals variance plus the square of the bias and that the rational
expectations character of the Bayesian Nash equilibrium eliminates all unconditionalbias from the middleman's
interpretationof the agent's signal. Solving for the communicationequilibriumbetween agent and intermediary,we
find39
2(kb) 1 + (kb)2(N2 _ 1)
where N = N(kb) is the largest numberof partitionelements given their dissonanceb - b! = kb. Substituting(A. 10)
in (A.9), we can rewritethe expected utility of the principalas
1 k2b2(N2 _ 1)
EUp(k, b) =Up(m, m) 1 (A.31) (1 - k) b
where N is the numberof partitionelements in the most informativecommunicationequilibriumbetween agent and

intermediary.FromLemma 1, N is given by
N = N(kb)_--2 + -2 + k , (A.12)
where (z) denotes the smallest integer greaterthan or equal to z. Let us denote by b [I - k*(b)]b the bias of the
intermediarywhich maximizes (A. 1), and by N* =- N(k*b) the numberof partitionelements in the corresponding
communicationequilibrium.Define furtherk(q) as
k(q) = 32
39. See CS, Section 4, for computations.

(A) We firstshow thatfor b e] 6, [, k* = k(q), where
q 1( 36b2+7+ 4j/636b2-1/3))
(2 2 ( 36b2- )1
and q is also the numberof partitionsin the correspondingcommunicationequilibriumbetween intermediaryand agent:

Supposethat N* = q and that given k = k(q), an equilibriumwith q partitionelements exists, i.e. N(k(q)b) > q.
From (A.11), then k* = k(q), which yields an expectedutility of
1 q2- 1b2
eup(q, b) - Up(m,m)- - q2 b2. (A.13)
12q2 q2
_ 2
Moreover,even if N(k(q)b) < q, we always have that
Vb: EUp(k*(b), b) < eu p (N*(k*(b)b), b). (A.14)
Define now the correspondenceb(q):
b e b(q) X eup(q, b) = eup(q - 1, b). (A.15)

Thus b e b(q) is a bias for which the principalis indifferentbetween delegatingauthorityto an intermediarywith bias
[1 - k(q)]b and bias [1 - k(q - l)]b, hereby assuming that a communicationequilibriumwith q, respectivelyq - 1
partitionelements then effectively exists. Substituting(A.13) in (A.15) and imposing b > 0, we find that b(q) is a
singleton;abusingnotation,we have
1 + - +
= (q /(q2 2)(q2 2q 3)
/
6 (q - 1)q
Moreover, from (A.13), if b = b(q), then for b > b(q), eup(q - 1,b) > eup(q,b), while for b < b(q),
eup(q - 1, b) < eu p(q, b). Since b(q) is strictlydecreasingin q, the following lemma thus holds:
Lemma A4. Ifb e]b(q + 1), b(q)[, then Vn e AN,n 7 q : eup(q, b) > eup(n, b).
If for b e [b(q + 1), b(q)], a communicationequilibriumwith q partitionelements exists, thatis if
3
Vb E [b(q + 1), b(q)] : q < N 2 -b) (A.16)
/ 3 b
then from LemmaA4 and inequality(A. 14), k* = k(q)= 2 .

From(A.12) and given thatq is an integer,(A.16) holds if and only if
11 / 2(q2 + 2)
q< + - + (q (A.17)
2 2 3b(q)
Substitutingthe value of b(q), which tends to 1/6 if q becomes very large, one can verify that inequality (A.17) is
satisfiedfor any q E AK.
LemmaA5. VqE ,q >l :b ]b(q + 1),b(q)[ == k* = k(q).
= X and limq,oo b(q) = 6. Invertingb(q) on ] ,

The function b(q) is decreasingin q, where b(2) --2] and
taking into account that q must be an integer,it follows from Lemma A5 that for b e ]1, _ [, k* = k(q) where q is
given by
/ 1 1 36b2 + 7 + 46,/36b2 -1/3\

\2 2X 36b2_ 1/
Note finally that we necessarilyhave that the numberof partitionelements in the resultingcommunicationgame cannot
be largerthan this q. Indeed, suppose N*(k*b) > q. Since for a given b and bj, the principalis always betteroff by
moreinformativecommunicationbetweenagentandintermediary,EUp (k*, b) > eup (q, b). However,from(A. 14) also
eup(N*(k*b), b) > EUp(k*, b). Since, by construction,b e [b(q + 1), b(q)], this in contradictionwith Lemma A4.
Hence, N(k*b) = q.
(B) To conclude the proof, we show that (i) for b < 1/6, k* = 0 and (ii) for b > _, k* = 1:
(i) One can easily verify thatfor b < 1/6, 1 + b2 > b2 as long as q is a finite number.Hence, from (A. 13)
12q2 q2+2
and (A.14), the principaloptimallydelegates to full authorityto the agent (or k* = 0) for b < 1/6 in which case
she obtainsa utility -b2.
(ii) One can easily verify that for b > /2/4, 1- + 21b2> 1/12 for any q e A. Hence, from (A.13) and
(A.14), the principaloptimallykeeps authority(or k* = 1) for b > /2/4, in which case she obtainsan expected
utility of - 1/12.
A.3. Delegation with veto-power or "closed rule"
Proof of Lemma 4. The proof for yo < L - 3b follows directlyfrom the proof of proposition8 in Krishnaand
Morgan(2000).40 While Krishnaand Morgando not considerthe case where Yo > L - 3b, one can easily verify thatas
long as yo < L - 2b, their equilibriumstill exists and is Pareto-dominant.In contrast,for Yo > L - 2b, the principal
would veto a proposalof y = Yo+ 2b if, as in KrishnaandMorgan,the agent were to proposethis wheneverm E [Yo,L].
The Pareto-dominantequilibriumis then economically equivalentto the one describedin Lemma4, (ii).
Proof of Proposition 6. As arguedin the text, for b < 1/6, the averagedeviationfrom m underthe closed rule
equals b. Due to the variancein this deviation,the principalthen strictlypreferspuredelegation.If 1/6 < b < 1/4, we
know from Lemma4, thatthe expected utility underthe closed rule is given by
1 b2dx + f x2 + dx x)x2dx+ = - b3b + .
If 1/4 < b < 1/2, it is given by
b 2d + x2dx
x x2dx
+ = b2 b3 + 24
The principal'sutility underpure delegation is given by -b2. It follows that pure delegation is strictly preferredover
delegationwith veto-powerif and only if b < 1/4.
Proof of Proposition 7. From Lemma 4, following the same reasoning as for Proposition 6, the average
deviation from m under the closed rule equals b as long as b < (L - yo)/3. The principalthen strictly prefers pure
delegationsince in the lattercase the deviationis constant.In contrast,for b > (L - yo)/3, the averagedeviationfrom
m underthe closed rule is smallerthan b. Hence, if the principalwere risk-neutral,she would strictly preferthe closed
rule over delegation.It follows that 3i(x) such that the closed rule dominatespure delegationif the loss functionof the
principal,e(x), is less convex in the sense of Arrow-Prattthant(x).
Acknowledgements. This paper is based on chapterthree of my Ph.D. dissertationsubmittedto the Universit6

Libre de Bruxelles and the Universit6de Toulouse. The authoris very grateful to Mathias Dewatripont,PatrickRey
and Jean Tirole for their guidance.I furtherthankPhilippe Aghion, Abhijit Banerjee,Rob Gertner,Oliver Hart,Patrick
Legros, Meg MeyerandMarcoOttavianifor very helpful discussionsandremarks,as well as seminarparticipantsat MIT,
LSE, Leuven,Chicago, Stanford,LBS, UniversityCollege London,Toulouse,Tel Aviv,Tilburg,Helsinki,Wisconsin,the
CEPR Conferenceon InformationProcessing Organizationsin Brussels and the SITE conferenceon contracttheory in
Stanford.He also thanksMarkArmstrong(the editor)and threeanonymousrefereesfor severalcommentswhich greatly
improvedthe paper.Financialsupportfrom the Foundationfor Scientific ResearchFlandersis gratefullyacknowledged.
REFERENCES
AGHION,P. and TIROLE,J. (1997), "Formaland Real Authorityin Organizations"Journalof Political Economy,105,
1-29.
ARMSTRONG,M. (1994), "Delegationand Discretion"(Mimeo).
AVERY, C. and MEYER, M. (2000), "Designing Hiring and Evaluation Procedures when Evaluatorsare Biased"
(Mimeo).
40. As argued by Krishna and Morgan, this equilibriumPareto-dominatesthe one proposed in Gilligan and
Krehbiel(1987).
BAKER, G., GIBBONS, R. and MURPHY, K. (1999), "InformalAuthority in Organizations", Journal of Law,
Economicsand Organizations,15, 56-73.
BANERJEE,A. and SOMANATHAN,R. (2001), "A Simple Model of Voice",QuarterlyJournal of Economics, 116,
189-228.
BATTAGLINI,M. (2001), "MultipleReferralsand MultidimensionalCheapTalk",Econometrica(forthcoming).
BERTRAND,M. and MULLAINATHAN,S. (1999), "Is There a Discretion in Wage Setting? A Test Using Takeover
Legislation",RandJournalof Economics,30, 535-554.
BRICKLEY,J., SMITH, C. and ZIMMERMAN,J. (1997) Managerial Economics and OrganizationalArchitecture
(Chicago:Irwin).
BOLTON, P. and DEWATRIPONT,M. (1994), "The Firm as a CommunicationNetwork",Quarterly Journal of
Economics,109, 809-839.
CRAWFORD,V. and SOBEL,J. (1982), "StrategicInformationTransmission", Econometrica,50, 1431-1452.
DE GARIDEL-THORON,T. and OTTAVIANI,M. (2000), "TheEconomics of Advice"(Mimeo).
GILLIGAN,T. and KREHBIEL,K. (1987), "CollectiveDecision-Makingand StandingCommittees:An Informationale
Rationalefor RestrictiveAmendmentProcedures",Journalof Law,Economics,and Organization,3, 287-335.
GROSSMAN, S. and HART, 0. (1986), "The Costs and Benefits of Ownership:A Theory of Vertical and Lateral
Journalof Political Economy,94, 691-719.
Integration",
HART, 0. and MOORE, J. (1990), "PropertyRights and the Nature of the Firm'",Journal of Political Economy, 98,
1119-1158.
HOLMSTROM,B. (1977), "OnIncentivesand Controlin Organizations"(Ph.D. Dissertation,StanfordUniversity).
HOLMSTROM,B. (1984), "On the Theory of Delegation",in M. Boyer and R. Kihlstrom(eds.) Bayesian Models in
EconomicTheory(New York:North-Holland).
JENSEN,M. (1986), "AgencyCosts of Free CashFlow, CorporateFinance,andTakeovers", AmericanEconomicReview,
76, 323-329.
KRISHNA, V. and MORGAN, J. (2000), "AsymmetricInformation and Legislative Rules: Some Amendments",
AmericanPolitical Science Review(forthcoming).
KRISHNA,V. and MORGAN,J. (2001), "AModel of Expertise",QuarterlyJournalof Economics, 116, 747-775.
LERNER, J. and TIROLE, J. (2000), "The Simple Economics of Open Source", Journal of Industrial
Economics(forthcoming).
LEGROS,P. and NEWMAN, A. (2000), "Competingfor Ownership"(Mimeo).
MITUSCH,K. and STRAUSZ,R. (1999), "Mediatorsand MechanismDesign: Why Firms Hire Consultants"(Mimeo,
Free Universityof Berlin).
MELUMAD,N. and SHIBANO,T. (1991), "Communicationin Settingswith No Transfers", RandJournalof Economics,
22, 173-198.
OTTAVIANI,M. and SORENSEN,P. (1999), "ProfessionalAdvice"(Mimeo).
RADNER, R. (1993), "TheOrganizationof DecentralizedInformationProcessing",Econometrica,61, 1109-1146.
RAJAN,R. andZINGALES,L. (1998), "Powerin a Theoryof the Firm",QuarterlyJournalof Economics,113, 387-432.
SEIDMANN, D. and WINTER, E. (1997), "Strategic Information Transmission with Verifiable Messages",
Econometrica,65, 163-169.
SPECTOR,D. (2000), "PureCommunicationBetween Agents with Close Preferences",EconomicLetters,66, 171-178.
STOCKEN,P. and MORGAN,J. (2000), "AnAnalysis of Stock Recommendations" (Mimeo).
SZALAY,D. (2000), "OptimalDelegation"(Mimeo).
TIROLE,J. (1999), "IncompleteContracts:WhereDo We Stand",Econometrica,67, 741-782.
VANZANDT, T. (1999), "Real-TimeDecentralizedInformationProcessingas a Model of Organizationswith Boundedly
RationalAgents",Reviewof EconomicStudies,66, 633-658.
VICKERS,J. (1985), "Delegationand the Theoryof the Firm",EconomicJournal(ConferencePapers),95, 138-147.

The Review of Economic Studies LTD

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Review of Economic Studies LTD

Uploaded by

Copyright:

Available Formats

The Review of Economic Studies Ltd.

Authority and Communication in Organizations

This paperstudies delegation as an alternativeto communication.We show that a principalprefers

1. Among firms decentralizingdecision rights in the 1990's are AT&T,GeneralElectric, EastmanKodak,Fiat,

CS-equilibrium,the finerthe equilibriumpartitionbecomes in absoluteterms,the more coarse it

The informational theory of legislative rules. Both the incentivebasedrationaleandour

de Garidel-Thoronand Ottaviani(2000), which also comparesdelegationwith communication,

UA(y, m) = UA(m+ b, m) - A(IY - (m + b))

their personnelmay fire too few employees duringa restructuring.Anecdotalevidence of other

Authority and contracts. We adoptan incompletecontractingapproachby positing that

As shown by CS, all equilibriain this communicationgame are characterizedby a partition

y(a, ) -- argmax f Up(y, m)f(m)dm.

Proposition 1 (Crawford and Sobel: Communication Equilibrium). If b > 0, then

UA(y(ai, ai+1), ai) - UA(y(ai- 1, ai), ai) = 0, (A)

Proof. The proof follows directlyfrom CS, Theorem 1. 11

Equilibrium selection. Although there is, in general, a multiplicity of economically

Comparative statics. Underthe same conditionsas above, CS establisha sense in which

A first comparison. From the previous analysis, the comparativestatics of delegation

5. THE OPTIMALALLOCATIONOF AUTHORITY

Key to our results will be the observationthatpartition elements are increasinglylarge as

Lemma 1 explains us neatly how the "screening mechanism" of the communication

The following corollaryis an immediateconsequenceof Proposition2:

Corollary1 providesa firstdeterminantfor the allocationof authority:delegationis optimal

The leading example of Crawford-Sobel. An interestingimplication of the previous

a closed-formsolution.Suppose F(m) is uniformlydistributedon [0, 1], Up (y, m) = -(y-m)2

Proposition 3. Consider the most informativecommunicationequilibriumgiven b. For

We providea sketchof the argument,a formalproof is given in the Appendix:

Lemma 2. If F(m) is symmetric,then communicationis informative(N(b) > 2) if and

Proof. See the Appendix. II

Proposition 4. Assume F(m) is symmetric and b < L/2, then informative

f g(lml)dF(m) < g(b). (9)

Proof. If (9) holds, the principalobtainsa higherutility by takingan uninformeddecision

Corollary 2. Assume F(m) symmetricand b < L/2. If theprincipal has a quadraticloss

Normal distributions. How largeis the parameterrangefor which delegationis optimal?

6. THE VALUEOF AN INTERMEDIARY

Proposition 5. If F(m) is uniformlydistributedon [O,1] and theprincipal has quadratic

* For b < 1/6, the principal delegates authorityto the agent.

Proof. See the Appendix. 11

Figure 3 summarizesthe optimalallocationof authoritywith and withoutan intermediary:

Lemma 3. If F(m) is uniformlydistributedon [-L, L] the optimalamountof discretion

Proof See Holmstr6m(1977, Section 2.3.2). 11

Lemma 4. Underdelegation with veto-power(closed rule):

(i) If b < (L- yo)/2:

* y= yO+2bifm e (yo, y +2b].

(ii) If b > (L- yo)/2:

Proof See the Appendix. 11

We comparethe outcome underpure delegationwith the above equilibrium.Given that the

Proposition 6. Consider the leading example of CS, in which F(m) is uniformly

Proof. See the Appendix. 11

(i) If b < b(yo), pure delegation strictlydominatesdelegationwith veto-power(closed rule).

Proof. See the Appendix. 11

8.1. Extendingthe space of contracts

B'(y)=KA-kA m+ (kA sk)b-

8.2. Uncertaintyover the bias and repeatedinteraction

de Garidel-ThoronandOttaviani(2000), if F(m) is uniformlydistributedandthe bias takeseither

8.3. Hard vs. soft information

31. Concretely,thereexists an equilibriumin which the principalrubberstampsall proposalsy e [-L + b, L -b],

Proof of Proposition 3. Let (ao, al,..., aN(b)) characterizethe most informativepartitionequilibriumgiven

Lemma Al. As h tends to zero,

wherewe have neglectedall terms in the 3-rd or higherorderof hn and hn+ .

(Im -Yn+l ) = 2t"(0)(m - Yn+l)2

where both m - Yn+l I and m - an are fractionsof hn+1.